ARTEMIS: After Removing No-Arbitrage, Directional Accuracy Falls Back to Coin Flipping

A number to open with: 64.89%.

That is the success rate of a model at judging price direction during crash regimes. Not spectacular, but significantly better than guessing. Then the researchers did something very plain: they deleted the no-arbitrage constraint from the training objective and changed nothing else. Directional accuracy dropped to 50.32%.

Fifty point three two percent is a coin flip. One physical constraint was the entire distance between "somewhat useful" and "useless" for this machine.

The paper is called ARTEMIS, posted to arXiv in March 2026, a single-author work by Rahul D Ray. The name expands to "Arbitrage-free Representations via Tractable Economic Models and Interpretable Symbols." It belongs to an emerging family: neuro-symbolic models for financial forecasting. Neural networks handle fitting, symbolic rules handle explanation, economics provides constraints. This material comes from an ApexQuant video breakdown; all numbers were checked one by one against the paper and all matched—but parts the video didn't cover are filled in below.

Three components, each with one job

First: timestamps are not noise. Financial data is extremely irregular—dense at open, vacant during lunch breaks, a full hour's order flow flooding in within a second during a crash. The traditional approach interpolates and aligns everything into a neat grid so models can process it. ARTEMIS's first step stops doing that: it uses a continuous-time encoder called the Laplace Neural Operator to consume irregular time points directly. This tool comes from the operator-learning lineage, decomposing inputs into a set of exponential bases via the Laplace transform, naturally suited to long-range memory—and "when something happened" is itself part of the input. The intervals between trade arrivals carry clues about market state; smoothing them away means throwing information out.

Second: teach economics as physics. In recent years, physics-informed neural networks work like this: write physical laws (heat conduction, fluid dynamics) as PDE residuals, stuff them into the loss function, and the network dares not learn anything that violates physics. ARTEMIS does the same thing, only swapping physics for finance's fundamental law: no-arbitrage. The no-arbitrage condition becomes a computable residual via the Feynman-Kac formula—if the model's price process implies "one dollar today guaranteed becomes two tomorrow," a free lunch, the loss function immediately punishes it. Another constraint governs the Sharpe ratio: if the strategy the model learns shows implausibly high instantaneous Sharpe ratios, it gets suppressed too. In finance, things that are "too good to be true" are usually not opportunities—they are overfitting.

Third: let the rules grow out of training. This is the most interesting layer. The model adds a symbolic bottleneck in the middle: latent variables are first distilled into sparse mathematical expressions—a handful of closed-form formulas built from few operators and coefficients—and then passed to the output layer. The whole process is differentiable and trained end-to-end. The effect is twofold: the bottleneck forces information compression, preventing overfitting; and the final model comes with a human-readable list of rules. Interpretability isn't a post-mortem performed by calling in SHAP after training—it's the architecture speaking plainly from the start. Post-hoc attribution and priori rules are two different roads: the former explains why a black box went wrong; the latter constrains how it can be right from the beginning.

The full scorecard, including the two losses

The video title claims "interpretable and SOTA simultaneously." That claim needs all four exams laid out.

On the DSLOB dataset: 64.96%, first place overall. On Time-IMM: 96.0%, also first, plus lowest error. Those two wins are clean. But the remaining two deserve to be stated: on Jane Street real trading data, directional accuracy was 51.5%, tied with LSTM—not a win. On Optiver volatility forecasting it was worse: directional accuracy of 45.82%, a low ranking, and a negative rank correlation. And of the four datasets, DSLOB is the paper's own synthetic data (a testing ground amplifying historical crashes into extreme regimes), while Time-IMM is temperature data. In other words: the two most eye-catching leads came one on synthetic data and one outside financial markets; on the two real financial exams, it tied once and trailed once.

The paper's own attribution is honest: the Optiver series is too long, the target is volatility, and features are too few—exposing the framework's applicability boundary. That boundary is worth magnifying: economic constraints help with the task of "direction judgment"; when the task becomes predicting volatility magnitude, the no-arbitrage guardrail is no longer on the road.

The ablation table hides a deliberate trade

Laying out the ablation experiments completely tells a story of four steps. Replace the whole model with an MLP: 35.04%, worse than guessing. Remove both physics losses at once: 41.77%, also worse than random. Remove only no-arbitrage: 50.32%. Remove only the Sharpe constraint: 56.82%. Keep everything: 64.89%. Each step prices an individual component, and the no-arbitrage step is the steepest.

There's something more subtle too. Removing the stochastic differential equation group causes error metrics to plunge—numerical accuracy of predictions improves dramatically while directional accuracy only slightly declines. It looks like "deleted a bad component," but the paper reads it as design intent: the stochastic term adds variance to predictions, sacrificing some precision in exchange for robustness in out-of-distribution regimes. Crash-regime stress tests support this—the SDE-equipped version degrades more slowly as conditions grow more extreme. This is the most honest passage in the paper: not every component improves every metric, and the authors know what they are trading for what.

Back to that number

The skeleton here is actually old: turning a field's first principles into loss functions is a move physics has long mastered. What's new is three things happening at once—the no-arbitrage constraint from finance entering a deep model's temporal path in differentiable form for the first time; irregular timestamps upgraded from dirty data awaiting cleaning to an input feature; and interpretability transformed from a post-hoc report into an architectural property.

For model builders, it offers a replicable ablation template: every domain constraint you add should be able to answer "how many points do we drop without it." For model watchers, the jump from 64.89 to 50.32 is itself a lesson: constraints are not decoration—they are the load-bearing wall of predictive power.

As for live trading—the paper uses only public and synthetic data throughout, and the author explicitly states live performance may differ significantly. The video's closing line is worth keeping: talk risk control first, returns second.

---

*References: Ray, R. D. (2026). ARTEMIS: A Neuro-Symbolic Framework for Economically Constrained Market Dynamics. arXiv:2603.18107; Berto, A. et al. Laplace Neural Operator for Solving Differential Equations (LNO lineage); video source: ApexQuant's ARTEMIS deep-dive, numbers verified against the paper.*